Fighting Voice Spam with a Virtual Assistant Prototype

Page created by Glen Moss
 
CONTINUE READING
Fighting Voice Spam with a Virtual Assistant Prototype
Fighting
                                            Voice Spam with a Virtual Assistant Prototype
                                            Sharbani Pandit1, Jienan Liu2, Roberto Perdisci1,2, and Mustaque Ahamad1
                                                                    1
                                                                        Georgia Institute of Technology
                                                                            pandit@gatech.edu
                                                                          mustaq@cc.gatech.edu
                                                                          2
                                                                            University of Georgia
arXiv:2008.03554v1 [cs.CR] 8 Aug 2020

                                                                          jienan.liu25@uga.edu
                                                                           perdisci@cs.uga.edu

                                               Abstract. Mass robocalls affect millions of people on a daily basis. Unfor-
                                               tunately, most current defenses against robocalls rely on phone blocklists
                                               and are ineffective against caller ID spoofing. To enable the detection of
                                               spoofed robocalls, we propose a virtual assistant application that could be
                                               integrated on smartphones to automatically vet incoming calls. Similar to
                                               a human assistant, the virtual assistant can pick up an incoming call and
                                               screen it without user interruption to determine if the call is unwanted. Via
                                               a user study, we show that our virtual assistant is able to preserve the user
                                               experience of a typical phone call. At the same time, we show that our system
                                               can detect mass robocalls without negatively impacting legitimate callers.

                                        1    Introduction

                                        Telephony has been a relatively secure channel for voice communication for over
                                        140 years. Currently, about 4.77 billion people across the world rely on the global
                                        telephony system for voice communication. Recent technological advances (e.g., Voice
                                        over IP) and policy changes offer many benefits, including lower cost and richer phone
                                        call experience. However, due to its convergence with the Internet, the phone channel
                                        has become a target of attacks by fraudsters and cyber criminals [1,2]. Cheap mass
                                        robocalling [5], voice phishing [3] and caller ID spoofing [4] are some of the techniques
                                        that are being used by fraudsters in these attacks. The number of robocalls continues
                                        to hit new highs each year. The anti-robocalling company YouMail estimates that
                                        February 2020 saw 4.8 billion robocalls [44] in the United States. Other companies
                                        have recorded similar high volumes of robocalls as well [18], and complaints about
                                        unwanted calls in August 2019 to the Federal Trade Commission (FTC) in USA
                                        numbered about 428,000 [13].
                                            In response to the increasing number of unwanted phone calls, a number of call
                                        blocking smartphone applications have appeared such as Hiya, Truecaller, YouMail etc.
                                        which are used by millions of users. Although such call blocking apps can be somewhat
                                        effective against scam calls, their effectiveness can be easily degraded with caller ID
                                        spoofing [21]. Such spoofing is easy to achieve, and robocallers have resorted to tricks
                                        like neighbor spoofing (caller ID is similar to the targeted phone number) to overcome
                                        call blocking and to increase the likelihood that the targeted user will pick up the call.
                                        To help reduce caller ID spoofing, both industry groups and regulatory bodies have ex-
                                        plored stronger authentication for call sources. Although the recently published IETF
Fighting Voice Spam with a Virtual Assistant Prototype
2       S. Pandit et al.

RFC 8588 describes SHAKEN/STIR approach to enhance trust in the source of a call
with signatures [31,29,30], its widespread deployment by carriers will likely take many
years. Furthermore, elimination of caller ID spoofing will not make all unwanted calls
go away, as phone numbers can be cheaply acquired and used to overcome blacklists.
    At a high level, the robocall problem resembles the email spam problem, in which
information about the source of an email could potentially be spoofed. Over the years,
the security community has been successful in developing effective spam filtering
solutions to ensure that email remains a viable means of communication [50,49].
However, most techniques used in such solutions cannot be applied to detect and filter
voice spam. This is because when a user receives a phone call, the only information
the user can rely on before answering is the caller ID (i.e., the calling phone number).
Namely, the context of the call (i.e., the content of the caller’s message) cannot be
used to detect if the call is spam or not.
     To detect unwanted robocalls and to provide the user with more meaningful call
context when a phone rings, compared to only relying on the (spoofable) caller ID,
we introduce RobocallGuard, a natural voice interaction model which is mediated
by a Virtual Assistant (VA). The VA mimics a human call screener (e.g., a secretary)
who picks up an incoming phone call and makes the user aware of the call only when
it confirms that the call is not a robocall or other type of spam. When a call arrives,
if the caller ID is not among the user’s contact list, the VA transparently picks up the
call and briefly interacts with the caller to determine if its source is a robocaller. Such
interaction aims to be natural for legitimate callers, while enabling the detection of
robocall sources who indiscriminately target a large number of victims. Furthermore,
such interaction with the VA enables learning the context of the call. Calls that are
not detected as spam are passed on to the user, and the context extracted from the
conversation between the VA and the caller is provided simultaneously, allowing the
user to make an informed decision on whether the call is unwanted or legitimate.
    Recently, a number of automated caller engagement systems that attempt to
collect information about a call source have been proposed. The Call Screen feature
available on the latest versions of the Android phone app provides call context to the
user via a real time transcript of the call audio. When an incoming call arrives, the
user is prompted with three options: answer, decline and screen. If the screen option
is chosen, Google Assistant engages with the caller to collect audio and generate a
transcript of the ongoing call. However, users are notified (i.e., the phone rings) of all
incoming calls (including robocalls) and user intervention is needed to screen such calls.
Google Voice allows screening calls to hear the name of the caller before users answer
the call. In contrast, Robokiller [20], a smartphone application, features an Answer Bot
that detects spam calls by forwarding all incoming calls to a server, which accepts each
call and analyzes its audio to determine if the audio source is a recording. Once the call
is determined to come from a human, it is forwarded back to the user. In Robokiller, a
caller continues to hear rings while the call is picked up, analyzed and forwarded back
to the user, which could negatively impact legitimate callers. Also, the audio analysis
techniques used by Robokiller can be countered by a more sophisticated robocaller,
and unwanted calls originating from human callers, such as telemarketers or human
callers hired by scam campaigns like Tech support, IRS etc., cannot be stopped.
Fighting Voice Spam with a Virtual Assistant Prototype
Fighting Voice Spam with a Virtual Assistant Prototype          3

    To the best of our knowledge, no systematic usability and effectiveness studies
have been reported of either Robokiller or Google’s Call Screen. Our goal is to explore
an automated voice-based interaction approach that maintains both caller and callee
user experience, eliminates user interruption and stops unwanted calls even in the
presence of spoofed calls. We evaluate RobocallGuard’s detection capabilities with
a corpus of real robocalls and conduct a user study to evaluate its usability.
    Although it may not be possible to stop all unwanted calls, we believe more
secure communication via the telephony channel can be supported by an automated
call screening agent that can detect and block such calls without degrading user
experience. The voice interaction model investigated in this paper aims to help achieve
this goal and we provide a proof-of-concept demonstration that it could be easily
supported by current smartphones.
    In summary our paper makes the following contributions.
  – To the best of our knowledge, we are the first to evaluate a call screening virtual
     assistant that uses automated call handling and audio analysis to defend against
     robocalls and other types of spam calls, including those that evade blacklists with
     caller ID spoofing. In addition, transcription of call audio recorded by the virtual
     assistant is used to provide meaningful context about incoming calls to a user
     when the phone rings.
  – Our virtual assistant aims to provide a mechanism that is similar, albeit much
     less “sophisticated”, to having a human call screener who can pick up phone
     calls and only forward to the user those calls which are likely wanted and record
     messages for the calls that are most likely unwanted. As a result, users are not
     annoyed with continuous ringing from unwanted calls.
  – To demonstrate the ability of the virtual assistant to detect robocalls, we have
     developed a proof-of-concept smartphone app named RobocallGuard. To this end,
     we experimented with a corpus of 8,000 real robocalls collected by a large phone
     honeypot, and show that all of them can be detected and thus blocked.
  – In addition, our proof-of-concept app allowed us to conduct an institutional
     review board (IRB)-approved user study to assess the usability of our virtual
     assistant. The results of this study demonstrate that the natural experience of
     a typical phone call is preserved for both callers and receivers, while benefiting
     from the ability to detect robocalls and other potentially unwanted calls.

2    Related Work

Abuse in the telephony channel has grown considerably and high volume scams that
rely on this channel have proliferated in the recent years [9,12,10,17,19,14].
    It has been shown that phone blacklisting methods provided by smartphone apps
(e.g., Truecaller [36], Nomorobo [39], Youmail [37], Hiya [38] etc.) or telephone carriers
(e.g., “ call protect services offered by AT&T [40], Verizon [41] etc.) can be helpful.
However, these services typically rely on historical data such as user complaints or
honeypot-generated information [21,22,11], and their overall effectiveness tends to
be low due to caller ID spoofing [54].
Fighting Voice Spam with a Virtual Assistant Prototype
4       S. Pandit et al.

    A number of research papers have explored how caller ID spoofing can be de-
tected [4,6,7,8]. However, the applicability of these approaches for protecting users
from fraudulent calls has not been demonstrated. One approach to better inform
users about the sources of their calls is to limit or prevent caller ID spoofing, thus
improving the efficacy of blacklists. The Federal Communications Commission (FCC)
has mandated US telecom companies to start using SHAKEN/STIR by June 30,2021
[56]. SHAKEN/STIR [31,29] enables the callee to verify the correctness of the caller
ID. However,SHAKEN/STIR might not be as effective if most of the scammers are
operating outside of US, or victims continue to fall for scams that don’t use caller
ID spoofing [48].
    Phoneypot [16] demonstrated the feasibility of using a telephony honeypot to
augment abuse information. Other papers have explored vulnerabilities in telephony
systems, their exploitation and the motivation for telephony-based attacks [8,12].
In recent work that conducted a large scale user study [34], it was shown that a
significant fraction of users fall victim to telephone scams. [55] evaluates the impact
of user interface design elements on user decision-making for robocall blocking ap-
plications. In contrast, our work aims to handle incoming calls and provide a solution
to block unwanted calls from reaching recipients without user intervention, even in
the presence of caller ID spoofing.

3     System Design
3.1   System Overview
In this paper, we propose RobocallGuard, a virtual assistant (VA) based solution
that can help defend against unsolicited phone calls. We developed a Smartphone app
which can screen incoming calls without user interruption and intervention. The app
hosts a VA, which works as a human secretary and receives incoming calls on behalf of
the user. If the incoming call is from a whitelisted caller, the VA does not pick up the
call and immediately notifies the user by ringing the phone. A whitelist can be defined
by the user, and can include the user’s contact list and other allowed caller IDs (such
as a global whitelist which consists of public schools, hospitals etc.). On the other hand,
if the call is from a blacklisted caller, the VA blocks it and does not let the phone ring.
However, if the caller ID belongs to neither a whitelist nor a blacklist, the VA picks up
the call without ringing the phone and initiates a conversation with the caller to decide
whether this call should be brought to the attention of the user. To make this decision,
the VA presents the caller with a challenge which must be passed to reach the callee.
The challenge can be thought of as an audio captcha which verifies the legitimacy
of callers. To keep the caller experience natural, we experiment with a simple audio
captcha, the name of the callee. Upon picking up the call, the VA asks the caller to state
the name of the callee. If the caller says the correct name, the call is passed to the callee
by ringing the phone. The transcript of the conversation between the VA and the caller
is also shown on the phone screen to provide the callee with additional context. If the
caller can not pass the above mentioned challenge, the VA blocks the call and notifies
the user of the blocked call through a pop-up app notification. The VA also makes a
decision if the call is from an unwanted human caller or a robocaller (discussed in detail
Fighting Voice Spam with a Virtual Assistant Prototype
Fighting Voice Spam with a Virtual Assistant Prototype         5

in the later sections). Upon making this decision, the VA ends the call and stores the
audio recording and transcript of the call for the user’s convenience. Each audio record-
ing and transcript is appropriately labelled (unwanted human caller or robocaller) by
the VA. Since our proposed solution does not depend on the availability of a blacklist
of known robocallers, it can be effective even in the presence of caller ID spoofing.

3.2   Threat Model

3.2.1 In-scope Threats In this section we describe the scope of threats that our
virtual assistant is designed to protect against.
    Mass robocalls: Previous analysis shows that most of the robocall attacks that
took place recently are mass calls. Attackers architect several campaigns such as tech
support, IRS, free cruise and so on to reach a large number of phone users. Since
the goal is to get a large coverage at minimal cost, attackers of such spam campaigns
rarely target their victims individually. As a result, a simple audio challenge provided
by the VA in the beginning of the call can filter out mass robocalls. Callers that
cannot pass this challenge are not able to directly reach the callee.
    Mass unwanted live calls from human: Our defense mechanism not only pro-
tects against robocallers, but also from unwanted human callers such as telemarketers
and debt collectors who repeatedly try to reach people at wrong phone numbers.
Unless, the caller knows the name of the callee, the callee is not interrupted by the
call and is only notified asynchronously via a message that includes call context.
    Spoofed Calls: The use of caller ID spoofing has increased significantly in the
phone fraud eco-system. Neighbor spoofing is a common tool used by attackers these
days. In our system, all incoming calls are picked up by the VA first, and audio
analysis and speech recognition techniques are applied to prevent spoofed calls from
interrupting the callee.
    AI equipped attack: Attacks where the attacker is equipped with AI are not
common in the phone fraud eco-system. However, with the availability of tools like
Google Duplex [32], attackers can craft more sophisticated attacks where robocallers
make a natural conversation with the other party and bypass our proposed defense
mechanism. However, unless the attacker knows the name of the callee (which is not
the case in mass robocalling campaigns), their call will not be passed to the user.
The VA might mislabel an AI equipped robocaller as an unwanted human caller, but
will still be able to stop the unwanted call from reaching the user.

3.2.2 Out of scope Threats Our VA does not protect against the following
types of attacks. The attacks discussed below are currently not common, but may
emerge in the future in an attempt to defeat intelligent phone call defense tools such
as the one we propose in this paper.
    Targeted attack: Our VA only protects against mass spam/scam calls. If the
attackers obtain the callees name associated with a smart phone number through
leaked private data, they can evade the VA by saying the correct name. Currently,
such targeted attacks are rare, but the increase of leaked private information may
pose such threat in the future. We discuss this in detail in Section 6.
Fighting Voice Spam with a Virtual Assistant Prototype
6       S. Pandit et al.

   Landline calls: RobocallGuard only protects callees when they use a smartphone.
Hence, malicious actors making landline calls using public directories are out of scope.

3.2.3 Possible Evasions A robocaller might try to bypass our proposed defense
mechanism in the following ways.
    Common name attacks: A common name attack is where a robocaller plays
a prerecorded message of carefully chosen common names to bypass the VA. Since
we evaluated the VA where the challenge is the callee’s name, such an attack could
evade the VA if the user has an identical or similar name to any of the common
names used by the attacker. However, the challenge can be made more difficult for
the attacker by requiring both first and last names.
    Master key attack: Another possible evasion technique is crafting a keyword
that can evade a large set of names. This keyword works as a master key that might
fool the VA to accept the crafted keyword as the correct name.
3.3   Design Goals
In this section, we state the design goals that are required from a defense system
designed to combat unwanted calls. Such a system needs to perform content analysis
since relying only on caller ID is not sufficient to stop unwanted calls.
  – Add an extra layer of security between the caller and the recipient of the call.
     This facilitates that all calls from unknown phone numbers (i.e., phone numbers
     not stored in the recipient’s contact list) are passed through the VA, which filters
     out robocalls and unwanted human calls. The motivation behind this design goal
     to add a challenge to the caller before the callee’s phone rings and interrupts
     him/her. The challenge should be easy and natural enough for a legitimate caller
     to pass, but harder for mass robocallers to pass.
  – Provide additional information to the recipient about the content of the call, so
     that they can make an informed decision to pick up the call. Such information
     should be extracted from audio received from the caller before the call is picked
     up by the callee.
  – Preserve user experience in regards of latency and accuracy. The VA should not
     make the phone call experience too unnatural for the caller or callee.
  – Ensure privacy when an incoming call is handled by the VA. In other words, the
     VA should run locally and the call is not transferred elsewhere. This ensures that
     the conversation between the caller and the VA is protected.
In our implementation of the VA we are able to achieve the first three goals. We also
believe that ensuring privacy is feasible and we discuss it in detail in the later sections.

3.4   User Workflow
We consider a scenario in which smartphone users install RobocallGuard. All incoming
calls are sent to the VA. Hence, the phone does not immediately ring and notify the
callee of the incoming call (Fig: 1). The VA makes a preliminary decision based on
the caller ID of the incoming call. Whitelisted callers are immediately passed to the
callee and blacklisted caller are blocked. The following steps are followed to handle
calls from neither blacklisted or whitelisted callers.
Fighting Voice Spam with a Virtual Assistant Prototype                                                      7

                                      Incoming Call

                                                                                              Spam Detector (SD)
                              Call Interceptor (CI)                   Pass the
                                                                        call
                                                           Unknown     to SD     Response                Robocall
                   Whitelist               Blacklist                                                                                     Transcriber
                                                            Callers              Recognizer              Detector

                        directly to user
                         Forward call

                                                Hang up
                                                  call

                                                                                                         from human caller

                                                                                                                             from robocaller
                                                                                      Pass the call to

                                                                                                                             Save recording
                                                                                                           Save recording

                                                                                                                               and transcript
                                                                                                            and transcript
                                                                                        transcript
                                                                                        user with
                                                                                                    Blocked     Blocked
                                                                                                  Human caller Robocaller
                                                          Fig. 1: System Architecture

 1. The VA picks up the call, greets the caller and asks for the name of the person
    the caller is trying to reach. This enables the VA to check if the caller knows the
    callee, and if it does not, it is likely that this call is unwanted.
 2. If the caller says the correct name, the VA detects it and passes the call immedi-
    ately to the callee by notifying them of the incoming call. The correct name is set by
    the user when they install the app. The VA also provides the transcript of the con-
    versation it just had with the caller to the user. While passing the call, the VA asks
    the name of the caller as well and provides it to the callee along with the transcript.
 3. After a predetermined time t1 has elapsed from the initial greeting and the caller
    has not said the correct name, the VA asks for the name again and looks for
    caller interruption while playing this message. The intuition behind this is to
    differentiate robocallers from actual human callers. Current robocallers typically
    play a pre-recorded message and do not stop to make a conversation with the
    callee. Therefore, the callers who are not interrupted by the VA are labeled as
    robocallers. On the other hand, if the caller is silent during most of the time the
    VA is talking, we label them as potential unwanted human callers.
 4. After t2 seconds have passed from the initial greeting and the caller has not said
    the correct name, the VA hangs up the call and saves the entire audio recording
    of the conversation it had with the caller. The VA also saves a transcript of the
    audio. The audio recording and transcript is labeled as robocaller or human caller
    according to step 4. Finally, the callee is notified of this blocked call and provided
    with the audio recording and transcripts. This allows the user to make a decision
    if they want to call back the caller or not.

3.5   VA Architecture
In this section, we describe our system architecture which is independent of the
implementation environment. The underlying architecture of our virtual assistant
is designed to fulfill the goals mentioned in Section 3.3. To achieve our first goal of
enhanced security, the modules, Call Interceptor and Spam Detector, act as a middle
layer. They collect and analyze additional call information to enhance security against
unwanted calls. The communication flow between the system components ensure that
user experience is preserved. To ensure privacy, our envisioned system handles all
calls locally. This could be achieved by implementing the VA as part of the default
8        S. Pandit et al.

phone app, for example. In contrast, Robokiller [20] routes all incoming calls to a
central server, thus exposing possibly sensitive audio to a third-party. Figure 1 depicts
the main components that make up our VA.
    We envision our system to be embedded in the Phone app of a smartphone.
The call screening feature provided by Google Pixel [23] phones demonstrates the
feasibility of embedding a VA with the phone app to intercept and examine voice
from incoming phone calls. However, due to certain OS enforced restrictions, we
implement a proof-of-concept prototype instead of embedding the VA with the Phone
app. Such limitations and the choices we make to overcome them are discussed in
Section 4. In the following, we describe each component of the system architecture.

3.5.1 Call Interceptor All incoming calls are passed to the Call Interceptor(CI)
module. The main function of the CI module is interception of a call to acquire the
incoming audio stream, and injection of recorded voice messages by the VA into
the outgoing audio stream. The CI makes an initial decision based on the caller ID.
All calls from whitelisted phone numbers are passed to the user (callee in this case)
without further processing. A user has total control of the whitelist on his or her
phone and can decide phone numbers from which calls should be passed to them
directly without intervention from the VA. All calls from blacklisted phone numbers
are dropped and stopped from reaching the user. The blacklist is predefined as well;
however, it is designed to be dynamic to include newly appearing malicious phone
numbers [21]. Phone numbers which are not present in the whitelist or blacklist are
labeled as ”Unknown Callers”. Audio stream from all unknown callers are passed
to the Spam Detector module for further analysis.

3.5.2 Spam Detector This module analyzes the audio coming from an unknown
caller to make a decision about the nature of the incoming call.
3.5.2.1 Response Recognizer Incoming audio stream from calls originating from
unknown callers are passed from the Call Interceptor to the Response Recognizer(RR)
module. The RR module decides whether to pass the incoming call to the callee. If the
call is considered unwanted, it is handled by the VA and not passed to the user. How-
ever, since the notion of an unwanted call is different for each user, it is difficult to define
an unwanted call. Hence, we take a conservative approach: if the caller knows the name
of the callee, we label that call as wanted; conversely, when the caller does not know the
name of the callee, that call is labelled as unwanted. The intuition behind this approach
is that phone calls from a person who knows the callee are less likely to be unwanted.
    The user is allowed to set the name(s) that should be accepted as correct by the
VA. We refer to the name(s) set by the user during the installation of the app as
correct name(s). Users may set multiple correct names as a valid recipient of phone
calls coming to their device. After the installation, during a future phone call, the
caller is asked who they are trying to reach. After the call has been picked up, we
set a limit of 35 seconds (value of time limit t2) to allow the caller to say a correct
name(s). The value of t2 is set empirically, keeping in mind that the VA speaks for 15
seconds during the 35 second time limit, hence the the remaining time period should
be enough for the caller to provide meaningful context. Moreover, t2 can be tuned to
Fighting Voice Spam with a Virtual Assistant Prototype       9

match user experience and context details. If the caller says any of the correct names
at any point during this 35 second period, the RR module recognizes it and passes
the call to the user along with the transcript of the conversation between the VA and
the caller. However, if the caller does not say any of the correct names, the call is
deemed as unwanted and not passed to the user.
    The backbone of the RR module is a keyword spotting algorithm which can detect
the right keyword. In our scenario, the correct name(s) of the callee is the keyword.
There has been a lot of research on keyword spotting algorithms which are used
in many commercially available products such as Amazon Alexa, Okay Google in
Google products etc. Hence, we explored existing systems that can effectively detect
a keyword. However, for such a system to be usable in our VA, high accuracy with
limited training examples, an open source toolkit and a light enough system to run
on a mobile device is required.
    Since the users set the correct name by making audio recordings of them pro-
nouncing the names, it is not feasible to collect a large number of audio samples from
the users. In other words, the keyword spotting algorithm will have access to only
a few recordings of each name. However, the system should have a high true positive
rate, and a low false negative rate. Snowboy [24], CMU Pocketsphinx [25], Honk
[26] are all open source keyword spotting toolkits that are light enough to run on
a mobile device. Based on our experiments, we found CMU Pocketsphinx has lower
accuracy than Snowboy, when trained with names. Honk requires a larger number
of audio samples to train a keyword. On the contrary, Snowboy requires only 3 audio
recordings to train a keyword. Snowboy also supports multiple keyword models, thus
multiple names can be set as keywords. Hence, we chose Snowboy to recognize the
name. Because Snowboy does not connect to the Internet, it can ensure privacy, which
is one of our design goals. We treat Snowboy as a blackbox, which when provided
with 3 audio samples, creates a model to detect the keyword. We embedded the
downloaded trained model with the VA to recognize the correct name(s).

3.5.2.2 Robocall Detector The objective of this module is to determine whether the
caller is an actual human or a robocaller. When a caller cannot say the correct name,
their call is handled by the Robocall Detector(RD) module.
    As mentioned in the workflow Section 3.4, the RD is activated after t1 seconds.
We set t1 to 20 seconds as we want to give the caller enough time to say the correct
name. The analysis of the data we collected from our user study shows that the initial
20 seconds is a reasonable time for the callers to say the correct name even if they
have to repeat the name. After 20 seconds have passed from the initial greeting and
the caller has not said the correct name, the VA plays an audio message to interrupt
the caller. Let the duration of this audio played by the VA be t3 seconds (we set t3
to 5 seconds in our experiment). During these t3 seconds, the RD module checks if
there is silence from the caller’s side. We use Voice Activity Detection (VAD) [27] to
determine if the audio coming from the caller’s side contains voice or silence. If the
caller is silent for at least t3/2 seconds while the VA is playing the audio message,
we label the caller as an actual person. On the contrary, if the caller is silent for
less than t3/2 seconds, it is labelled as a robocaller. Determination of the type of
the caller (human or robocaller) provides additional information to the user about
10      S. Pandit et al.

the call. Upon determining the type of the caller, the VA allows the caller 10 more
seconds as a margin of error to say the correct name before hanging up the call. The
audio recording and transcripts of the entire conversation with the caller is saved
locally at the device for the user to preview later. The associated label (human or
robocaller) is used to determine in which folder the audio and transcript is stored.

3.5.2.3 Transcriber The transcriber component transcribes the entire conversation
between the VA and the caller to provide additional context to the user. The VA
stores the audio recording and the transcription of that audio recording locally; and
notifies the user of these two files after it has handled the incoming call. This helps the
user to access the content of the call without picking up and engaging in the phone
call. Calls that are passed to the user by the VA after deeming them as “wanted”,
are also provided with a transcript of the conversation between the caller and the
VA that took place before the call was forwarded to the user. When the user is
notified of a such a call by ringing the phone, the transcript is shown on the screen
for additional context. Calls which are deemed “unwanted”, both from human callers
and robocallers, are not passed to the user and hung up by the VA.
    The VA does not engage in a conversation with callers that are whitelisted and
passes the calls directly to the user. Therefore, transcripts are not provided for such
calls. All other callers are greeted by the VA and hence transcript is made available
to the user to understand the content of the calls.
    There are many software libraries and APIs available for transcription. We have
chosen Google Cloud Speech API [28] because it has a very high accuracy and tran-
scription can be performed from a mobile device, unlike Kaldi [42] and Mozilla deep
speech [43]. Ideally, the transcription should be conducted locally in the device and
no server should be involved. Android provides the means to perform transcription
locally through RecognizerIntent and SpeechRecognizer class. However it imposes
the restriction of the input channel. RecognizerIntent and SpeechRecognizer class
always uses audio from the microphone as an input for transcription. As discussed
in Section IV, audio from the caller in our implementation comes though the VoIP
channel instead of the microphone. Therefore, in our proof-of-concept prototype, we
do not perform transcription locally. Instead, we send the stored audio recording of
the conversation between the VA and the caller to Google Cloud and a correspond-
ing transcript is returned. However, when RobocallGuard is deployed in real life,
transcription can be done locally in the device, thus privacy can be ensured.

4    Implementation

In this section, we discuss some important details of our implementation of the VA.
We implemented a prototype of our app using Java on Android. We envision our
VA to be embedded with the Phone app where the VA handles all incoming calls
locally without having to stream audio to an external server. The recent release of
call screening feature for Android phones further supports the feasibility of such
a system. However current Android system restrictions do not allow embedding a
system that can inject voice messages in the outgoing audio stream without OS
Fighting Voice Spam with a Virtual Assistant Prototype         11

                                                       Response
                                Fig. 2: VoIP Architecture

modifications. Hence, in the following, we describe and explain the choices we made
in the implementation of our proof-of-concept prototype.

4.1   Motivation behind VoIP

The workflow of the VA starts with an incoming call being passed to the CI module.
As discussed earlier, the CI module captures the audio stream from caller side and
takes full control of the stream so that the system is capable of analyzing the voice
data, as well as locally recording the caller’s audio. Moreover, the CI also injects audio
into the phone call to communicate with the caller on behalf of call recipient, while the
recipient has no awareness of the incoming call during the time the VA is interacting
with the caller. To perform the first task, we could implement a customized phone
call app by making modifications to phone call service codebase provided by Android
(i.e. Implementing customized Android.telecom.InCallService). However, it is strictly
constrained for a common developer to satisfy all the requirements for injecting audio
to an ongoing phone call. For the sake of security and privacy, Android does not allow
injecting sound files in the conversation during a phone call [46], which means that
no such API is provided by the Android system that could pre-process or replace
the microphone as an input audio stream during a phone call.
     Taking all these into account, we decided to implement a VoIP(Voice over Internet
Protocol) application to conduct our user study experiment. With an VoIP phone
call application, we are able to get full access to voice streams on both sides of a
phone call. Furthermore, it is possible to inject audio at any appropriate moment
in the conversation during a phone call with a VoIP app.

4.2   VoIP application architecture

Our VoIP application consists of two major parts, namely client-side and server-side
systems. For the sake of convenience, we implemented the system so the two parts work
over TCP network connections by using a simplified and customized session initiation
and keep protocol. Figure 2 demonstrates the architecture of our VoIP application.
    On the server side, an application server (Server) working along with Google
Firebase Cloud Messaging Platform (FCM) [45], handles call session setup and main-
taintenance. On the client-side, we split the core of our call implementation into two
parts, one called “PhoneCallDialer (PCD)” and the other called “IncomingCallHan-
dler (ICH)”. PCD is mainly a dial service, which provides a dialing panel for the user
12      S. Pandit et al.

to make VoIP phone calls. At call session setup stage, PCD initiates a connection
request and builds a call session between caller and callee with the help of the Server.
At call session maintain stage, PCD streams and transmits caller’s voice to callee
through the channel maintained by Server.
    ICH consists of two functional modules, which implement callee-side handling
of incoming calls. The CI module automatically answers an incoming call without
the callee’s awareness during call session setup stage and makes the initial decision
to forward calls from whitelisted numbers or deny calls from blacklisted numbers.
For unknown callers, the call is passed to the Spam Detector (SD) module and
corresponding pre-recorded audio tracks are injected into the call to allow the VA
to communicate with the caller.
    We implemented and used the system described above to conduct a user study
and evaluate the efficacy of the VA. The main purpose of developing a VoIP prototype
is to assess the usability of the VA and its effectiveness in detecting robocalls and
other unwanted calls. Although we developed a VoIP prototype, it is possible to
embed our VA in the Phone app.

5     VA Evaluation

In this section, we report the results of experiments we conducted to measure the
accuracy of decisions made by our VA for incoming calls, and discuss a user study
that was conducted to evaluate the usability of our prototype.

5.1   Usability Study

Our VA is designed to provide the convenience of a human assistant while detecting
unwanted calls. It also provides context for calls, which helps the callee decide if a
call needs his/her attention. To explore the usability of such a system, we conducted
an Institutional Review Board (IRB) approved user study. In the following, we first
describe the study setup, its participants and then discuss the results.

5.1.1 Study Setup Our study participants consists of 21 users who were sampled
from a population of college students. Most of the participants can be described as tech-
savvy. All participants were required to be fluent in English and be familiar with using
smartphones. Each experiment was conducted with two Android devices, a Samsung
Galaxy S9 plus and a Samsung Galaxy Tab A, running Android 8.0 and 7.0 respectively.
Both of the devices had RobocallGuard app installed. During the user study, all phone
calls were made using the Tab and received using the smartphone. The setup of the
experiment is as follows. We briefed the participants about the experiment process and
explained the purpose of the VA. We asked the users to perform a list of tasks: making
calls, receiving calls, checking the contents of blocked calls and answering multiple
choice questions regarding their experience of using RobocallGuard. Participants were
assigned the role of a caller and a callee one at a time and were asked to make/receive
a call. When participants were assigned the role of a caller, they made 4 calls. Such
a call took at most one minute. When participants were assigned the role of a callee,
Fighting Voice Spam with a Virtual Assistant Prototype        13

they received 5 calls. After making/receiving each call, the participants had to answer
multiple choice questions regarding their experience. At the end of the user study,
each user was asked 6 generic questions about their overall experience with the app.

5.1.2 User Actions In this section, we describe each task the participants per-
formed during the user study in detail. Each experiment involved a pair of users
(user A and user B), one caller and one callee, performing the tasks. Once user A has
completed all the tasks assigned to the caller, they are assigned the role of a callee
and vice versa. The experiment starts with user A acting as the caller and making
calls to the callee, user B.
     We performed two sub-experiments within each experiment. During sub-experiment
1, we provided the caller with the appropiate response to the challenge i.e. the cor-
rect name. Hence, the caller should be able to reach the callee. Conversely, in
sub-experiment 2, the caller is either given an incorrect name or no name at all.
Therefore, the VA would not allow them to reach the callee. The participants had
no idea about what the correct name was. Furthermore, there are two scenarios in
each sub-experiment; one where the caller is provided with a script to read from
when making a conversation with the VA, and one where the caller is given a topic
to talk about, instead of a script, while interacting with the virtual assistant (e.g
calling a friend to make movie plans.) When a user is assigned the role of a callee,
with each forwarded call they are given the choice to either pick up or decline the call.
They are advised to use the caller-VA interaction transcript provided by the app to
make this decision. Once they pick up the call they can start a normal conversation
with the caller. During our user study, we preset the correct name to be Taylor
instead of having each user set a name. We make this choice because the purpose
of the user study is to get insights about call experience in the presence of a VA,
rather than testing the accuracy of the keyword spotting algorithm. We ask the
callee to impersonate Taylor when making the decision to answer an incoming call.
Furthermore, the callee has no advance knowledge of the content of the call. Following
is the detailed description of the 4 calls made by the caller during the experiment.
     Sub-experiment 1: In this case, the correct name “Taylor” is provided to the
caller and the callee is asked to impersonate Taylor.
     Scenario A: When asked the name of the callee, the caller is instructed to read the
following script, “Hello, can you please forward my call to Taylor?”. In this scenario
the VA allows the caller to reach the callee.
     Scenario B: In this scenario, the caller is not given a script. They are instructed
to make a call to their friend Taylor to make movie plans. They are advised to include
the name Taylor in their conversation.
     While forwarding the call to the callee, the VA asks the caller to state their name.
We advised our participants to say a fake name to protect their privacy.
     Sub-experiment 2: In this case, an incorrect or no name is provided to the
caller and similar to sub-experiment 1, the callee is asked to impersonate Taylor, the
correct name being set as Taylor.
14              S. Pandit et al.

                                                                                         15
  12

  10

                                                                                         10
     8

     6

                                                                                          5
     4

     2

                                                                                          0
     0                                                                                        Strongly       Disagree     Neutral    Agree        Strongly
         Strongly Disagree   Disagree         Neutral        Agree      Strongly Agree
                                                                                              Disagree                                             Agree
                Question 1       Question 2         Question 3       Question 4
                                                                                                         Question 5     Question 6   Question 7

         (a) Caller and Callee Response                                                       (b) User Generic Response

                                                                      Fig. 3: User Results

    Scenario A: In this scenario the caller is asked to read the following script, “Hello,
can you please forward my call to Robert? We met at a seminar today.” It should
be noted that Robert is the incorrect name here, hence the VA blocks this call.
    Scenario B: In this scenario, the caller is not given any script or name. They are
instructed to make a call to an office trying to sell a computer. Since the caller does
not say the correct name during this call, the VA blocks this call as well.
    As a result the callee gets a notification of the blocked call along with the transcript
of the call. After making each call, the caller is required to answer survey questions
that focus on the ease of interacting with the VA, the quality of the transition from
the VA to the callee and the delay experienced before the callee responded. As a
callee, a participant received the aforementioned 4 calls. The VA passes the first
two calls to the callee since the correct name was said by the caller. On the other
hand, the VA blocked the last two calls, notified the callee of the blocked calls and
provided her with the transcript and audio recording of the conversation. In addition
to receiving the 4 calls made by the caller, the participant received a robocall made
by us. The VA blocks this call, labels it as a robocall and notifies the user. After each
call (both blocked and passed), the callee is required to answer a number of survey
questions regarding the usefulness of the transcript, the interaction with the caller,
and the reason behind their decision to pick up or not pick up the call.

5.1.3 User Study Results In this section we present the results obtained from
the user responses to the survey questions. We present user responses to the following
seven questions in Figure: 3. Questions 1-2, 3-4, 5-7 investigate caller, callee and
overall user experience respectively.
  – Question 1: It was easy to interact with the VA.
  – Question 2: The delay you experienced before the other person responded to the
    call is acceptable.
  – Question 3: The transcript was able to provide sufficient information to infer the
    topic of the incoming calls.
  – Question 4: The transcript was able to provide sufficient information about the
    content of the blocked calls.
Fighting Voice Spam with a Virtual Assistant Prototype         15

 – Question 5: I found the app beneficial to me as it provides prior knowledge about
   the incoming calls.
 – Question 6: I think I would like to use an app equipped with a VA frequently.
 – Question 7: I felt comfortable with the VA intervening in the phone calls.
    Caller Experience: Most of the users who acted as the caller reported that it
was easy to interact with the VA. Figure 3a shows the distribution of the responses to
a question about the ease of interaction with the VA when the caller said the correct
name. When the caller was given an incorrect or no name (in case of sub-experiment
2), 4 out of 21 users disagreed that it was easy to interact with the VA. This is
understandable since the user had no knowledge of what the correct name is, they
became frustrated when the VA did not pass their call to the callee, which is what
may happen when a telemarketer calls. Figure 3a also shows the user responses
regarding the delay experienced by the caller before the callee could be reached. Most
users agreed or strongly agreed that the delay was acceptable. Moreover, all but one
caller reported that they were satisfied with the transition from the VA to the callee.
Hence, it can be concluded that a call experience for callers was not degraded as a
result of having the VA acting as an intermediary.
     Callee Experience: While acting as the callee, when a call is passed to the
user, they may either answer or decline the call. Most of the users who picked up the
call said that they made this decision because the transcript suggested this to be a
non-spam call. Only 4 users reported that they answered the call because they pick up
all incoming phone calls. All but one callee said that the interaction with the caller felt
normal/natural after they picked up the call. Figure 3a shows how the callees felt about
the transcript of the incoming calls. Only 3 users reported that the transcript was not
able to provide with sufficient information about the incoming calls. The reason behind
this is that some callers only said the correct name “Taylor” and nothing else after
starting a conversation with the VA. Thus, the transcript consists of only the correct
name and no other information. Since it is not possible to control what the caller
says, the quality of the transcript cannot be guaranteed in all cases. However, since
the caller is also required to say their own name before a call is forwarded, it provides
further information about the call to the callee. The current version of our prototype
does not ask the caller to state the purpose of the call. We plan to add this feature
in our future version, which we expect will augment the content of the transcript.
Figure 3a also shows callee responses about the transcript of the blocked calls. All
callees reported that they were notified in a timely manner of a new blocked call.
     Overall Experience: At the end of the user study, each user was asked 6 generic
questions. According to Figure 3b, most users mentioned that they found the app bene-
ficial to them as it provides prior knowledge about incoming calls. Moreover, most users
said that they would like to use an app equipped with a virtual assistant frequently. In
addition to these questions, all but two users said that the app was easy to use. Only
3 users reported that they needed to learn a lot of things before they could get going
with this app and only two users reported that they could not easily navigate through
the app. Moreover, all but 3 users reported that they were comfortable with the VA
intervening in their phone calls before the calls are forwarded to the callee. Based on
these results, we can conclude that the the VA does not negatively impact the overall
16         S. Pandit et al.

call experience and lack of usability will likely not be the reason for impeding its adop-
tion. Since we recruited 21 users for this study, the accuracy/quality of our conclusion
from this study lies between 80% to 95% (between 10 and 20 users) according to [53].

5.2      VA Performance
In this section, we evaluate the effectiveness of the VA based on its ability to forward
legitimate calls and block and label unwanted calls. For each incoming call, the VA
has to make a decision whether to forward the call to the user or not. Besides, it
needs to label each unwanted caller as human or robocaller.

5.2.1 Interaction with Human Callers We first measure the effectiveness of
the VA when the caller is a human. We categorize all human callers into two catagories:
legitimate callers and unwanted callers. We define a caller to be legitimate if they
know the correct name of the callee. On the contrary when the caller fails to pass
the challenge, we define them as unwanted callers. To measure the VA’s effectiveness
when callers are human, we used the data collected during the user study. We have
discussed usability insights in our user study results in Section 5.1.3; here we present
statistics on the VA’s accuracy.
    Legitimate callers: As discussed earlier, during the user study, we performed
four experiments with each of the 21 callers. In the first two experiments the callers
were given the correct name of the callee. Our analysis shows that when callers passed
the challenge, the VA detected all of them as legitimate callers and forwarded their
calls to the callee. In only two cases, the callers had to repeat the correct name, and
after the second time the callers had said the correct name, the VA forwarded the
calls to the callee. This was mostly due to snowboy, and advancements in keyword
spotting algorithms will further reduce such false negatives.
    Unwanted Callers: During the user study we conducted, in the third experiment
the callers were given an incorrect name and in the last one the callers were not given
any name at all. Therefore, in these cases, the callers in our user study played the
role of an unwanted caller in a way similar to a telemarketer, and made calls to the
callee. The VA picked up the calls and was able to prevent every incoming unwanted
call from directly reaching the callee.

5.2.2 Interaction with Robocallers In this experiment we measure the fraction
of robocalls stopped and correctly labelled by the VA. We use a dataset of phone
call records collected at a large phone honeypot provided by a commercial robocall
blocking company. This dataset contains 8081 calls (which had an average duration
of 32.3 seconds) coming into a telephony honeypot during April 23, 2018 and May
6, 2018 3. It records the source phone number that made the call, the time of the
call, the audio recording of the call and the transcript of the audio recording. Since
it is a telephony honeypot, it contains some misdialed calls along with robocalls.
     Robocall Detection: To filter out misdialed calls, we use the approach previ-
ously proposed by Pandit et al. in [21] to extract important topics from the transcripts
3
     Although this dataset is not recent, robocaller behavior has not significantly changed.
Fighting Voice Spam with a Virtual Assistant Prototype         17

of calls in our robocall dataset. Using LSI topic modeling, we extract 60 topics from
our corpus of transcripts where each topic represents a spam campaign. We construct
a similarity matrix by computing the cosine similarity between each transcript. We
then convert the similarity matrix into a distance matrix by inverting the elements
of the similarity matrix. We performed DBSCAN clustering on the distance matrix.
DBSCAN is one of the most common clustering algorithms which given a set of
points, groups together points that are closely packed together, marking as outliers
points that lie alone in low-density regions. At the end of this, 79 clusters were created
where each cluster represents groups of highly similar transcripts of robocalls. Since,
the honeypot by nature contains unwanted calls, the clustering technique acts as a
sieve that filters out all non-spam calls as outliers. Moreover, it allowed us to take one
representative robocall from each cluster and use the audio recording to make a call
to the VA. Upon making the calls, the VA correctly detected 100% of all robocalls
as unwanted and stopped them from ringing the user’s phone. Theoretically, a false
negative occurs when the VA forwards an unwanted call i.e. recognizes an incorrect
response to the challenge as a correct one. A false positive occurs when the VA blocks
a wanted call i.e. fails to detect the correct response to the challenge. With stricter
challenges and fast advances in AI, we expect the correctness of the VA to increase.
    Robocall Labeling: Once unwanted calls are detected, our VA determines if the
caller sounds like a human or a robocaller. From the previous experiment, we noticed
that the robocalls with a duration of less than 20 seconds were being inaccurately
labeled as unwanted human callers. The reason lies in the design of the VA. The
VA interrupts a caller at the 20th second and determines if they were interrupted
or not by looking for voice activity from the caller side while the VA is speaking. The
robocalls that are less than 20 seconds are found to be silent when the VA interrupts
and are mislabeled as human. To solve this problem, we analyzed the contents of the
short length robocalls. Since the robocallers are trying to financially profit from their
victims, the content of the short robocalls must serve a purpose. From analysis of the
robocall recordings in the honeypot, we discovered that 86% of the short robocallers
ask the callee to press or enter a digit in the phone keypad. Hence we take a further
step to identify the short robocalls. If the transcript of a call contains the keywords
”press” or ”enter”, we label it as a robocall. It is unlikely that a legitimate human
caller will say these words while interacting with the VA. With this added step, our
results show that the VA is able label 97.8% of all robocalls correctly.

5.3   Comparison with Call Blocking Apps
There are several commercial applications available in the app stores that aim at
blocking robocalls. Most of these apps (such as Youmail, Hiya, Nomorobo, etc.) rely
on phone blacklists. These blacklists are generated from user complaints of unwanted
calls (e.g., FTC reports) and other data collected by the app vendors. Thus, these
apps block an incoming call by looking at its caller ID and not the content of the
call. As discussed earlier, spoofed calls, which are common, cannot be blocked with
this approach. Currently, only Robokiller claims to perform call content analysis, in
addition to using phone blacklists, to block unwanted calls. We performed a small scale
experiment to compare Robokiller with RobocallGuard. To conduct this experiment,
18      S. Pandit et al.

we installed Robokiller on a Samsung Galaxy S9 Plus Android device. We then used
a Twilio [47] phone number to make 10 robocalls to the device where Robokiller is
installed. We chose a random sample of 10 robocall messages. Since the phone number
we used is a Twilio verified phone number, it is not a blacklisted phone number.
Therefore, a defense system that relies only on phone blacklists will not be able to
block these robocalls. Only a system that performs audio content analysis will be able
to detect these robocalls. However, we noticed that Robokiller was not able to block
any of the robocalls made from the Twilio phone number and let the calls pass to the
user without any spam call warning. In contrast, RobocallGuard was able to block all
of these 10 robocalls. Therefore, it can be inferred that the available Robokiller app
seems to rely mostly on phone blacklists rather than call content analysis. To explore
this further, we downloaded FTC user complaint reports from August 3-5, 2019 and
extracted 10 phone numbers that had most complaints. We then spoofed each of
these 10 caller IDs using SpoofCard [51] to make phone calls to Robokiller. Since we
conducted this experiment on September 10, it can be expected that the top 10 callers
from a month old FTC dataset will be in blacklists used by commercial apps. In
addition, we performed reverse look up on each of these 10 phone numbers and found
7 of them to be labeled as “scam or fraud”. Upon making the spoofed phone calls, we
found that Robokiller was able to block calls from 9 of the 10 caller IDs. This shows
that Robokiller is able to block most of the incoming calls from blacklisted caller IDs.
    In our next experiment, we downloaded FTC user complaint reports from Septem-
ber 9, 2019 and extracted the 10 phone numbers that were least complained about.
We spoofed each of these 10 caller IDs to make phone calls to Robokiller. Since this
dataset contained complaints from the previous day, it can be assumed that these
phone numbers are not present in blacklists. Upon making the spoofed calls, we found
out that Robokiller let calls from 8 of the 10 caller IDs pass. Although this is a small
scale experiment, it provides evidence that Robokiller appears to rely more on phone
blacklists and not the content of the call. Hence, unlike our proposed VA, robocalls
with spoofed or previously unseen caller IDs can evade Robokiller.

6    Discussion
RobocallGuard is the first version of our system to mainly test the usability of a VA
and its effectiveness in stopping current robocalls. The survey responses from the user
study we conducted show that both callers and callees are comfortable with the change
in the call experience due to the VA and found the VA beneficial. Moreover, [55] has
shown that users need a app which handles spam calls without making the phone
inoperable. The fact that RobocallGuard filters out spam call without user intervention
and without making the phone inoperable adds desired convenience to the users.
Much of the current voice abuse over telephony is perpetrated by mass robocallers
who indiscriminately call a large number of potential victims. They use techniques
like neighbor spoofing to increase the likelihood that their calls are picked up. Our
results show that RobocallGuard can be effective against such mass robocallers.
    We understand that the callee’s name might seem like a simple challenge, which
could be evaded by bad actors by obtaining names associated with phone numbers
from leaked data. However, most of the robocallers currently make cheap mass
You can also read